- Email: [email protected]
Soccer team based on agent-oriented programming
Robotics and Autonomous Systems Robotics and Autonomous Systems 21 (1997) 167-176
ELSEVIER
Soccer team based on agent-oriented programming J.L1. d e l a R o s a a,., A. O i l e r a, j. V e h i a, j. P u y o l b a Intelligent Systems and Control Engineering Group, Institute oflnformatics and Applications (HiA). University of Girona, Catalonia, Spain b Artificial Intelligence Research Institute (IliA), CSIC, Catalonia, Spain
Abstract
In this paper the analysis, design and implementation of a soccer team of micro-robots is explained. Besides the technical difficulties to develop these micro-robots, this paper also shows how to develop a multi-agent co-operative system by means of Matlab/Simulink (MATLAB and SIMULINK are Trade Marks of Math Works. Windows 95 is a Trade Mark of Microsoft Corporation.) a widely known computer aided control system design framework. Agent-oriented paradigms (AOPs) formalise interactions between multiple agents in terms of changing their mental states by communication between agents. Their practical implementations are usually conceived by means of object-oriented paradigms. Nevertheless, the implementation of agentoriented paradigms in Matlab/Simulink is not straightforward. Thus, the obtained real implementation is an integrated system that includes several programming paradigms so as hardware platforms. Finally, the proposal of the integrated framework for the micro-robots soccer team is shown. Keywords: Agents; CACSD; Soccer
1. Introduction
1.1. Mirosot
Multi-agent based mobile robotics require new examples from application and call for new control schemes. By this way, Micro-Robot World Cup Soccer Tournament (MIROSOT) [11] meeting is a forum where rules and constraints about the way that multiple micro-robots could work together in a game is a chance to apply techniques focused on agents. This game is the micro-robotic soccer where three micro-robots have to play against another team of the same characteristics and try to score as much as possible.
According to MIROSOT specifications, soccer micro-robots are limited to the size 7.5 × 7.5 x 7.5 cm, and must run autonomously on a 130 × 90 cm sized ground. MIROSOT rules have strong analogy with real soccer rules although adapted to current robotics technology. Teams can use a centralised vision system to provide robots with situations in terms of their position and orientation so as with the ball position. The vision-based facility was widely used by participants to the MIROSOT tournament. The vision-based measures are used by a computer host, here called HOST, to calculate individual robotic movements that are send (broadcasted by an FM emitter) to the robots to be executed. This is depicted in Fig. I. Thus, since robots have autonomous behaviour but the whole team
* Corresponding author. 0921-8890/97/$17.00 t997 Published by Elsevier Science B.V. PH S0921-8890(97)00024-9
J.Ll. de la Rosa et al./Robotics and Autonomous 5),stems 21 (1997) 167-176
t68
...',./:- !~ "~:..":. . . . . . . . . . . . . . .
i ..
VIDEO CAMERA CCD
........ "-..,.
R O B O T C...,OLORS ( T O P VIEW )
7.5 crn, I .: .~ : .,: ,i:
~ TEAMCOLOR (frontpart) .*---- PLAYER COLOR (bact~part)
Fig. I. The overall system. must actuate as a group, they must co-operate. This means that robots must communicate each other, and the proposal of this work is to deal with a more general problem, that is to do reasoning on communication as agents do. Therefore, agent-oriented paradigm (AOP) frameworks [13] are used in this work.
1.2. Agent-oriented paradigms and control Nowadays, a commonly proposed solution to the behaviour control in a community of robots is the application of agents theory [3]. The control of the global behaviour of a soccer team problem gives the opportunity to apply agents theory due to the distributed architecture of the mobile robots and the problem itself that implies coordination, competition, and cooperation by means of communication among each soccer player and perception. Therefore, the communication of information and knowledge is the most important action that might be taken by each agent, in this case, each soccer player. On the other hand, control engineering projects are developed as usual in the following steps: analysis, design, and implementation. There are modern computer-based tools that handle the control engineering practice, named computer aided control systems design (CACSD) environments, The robotics applications could be more easily developed by using such environments because of the availability of linear and non-linear systems analysis and design tools that could be useful to control robotics movements.
These two approaches are clearly different: the AOP provides us the conceptual model to describe, in a declarative style, the behaviour of the team by means of high level constructs and the current CACSD are not good at representing declarative knowledge but procedural. When dealing with complex systems further declaration is fundamental. Therefore, at the moment to solve complex systems control then the distribution of knowledge is essential in addition to automatic control, and the case of the soccer micro-robotics game is specially adequate. Therefore in this paper the conjunction of AOP and CACSD systems is proposed, because the AOP introduces declarative knowledge in terms of communication operators for dealing with decisions, capabilities and commitments of agents that are necessary in this project. But these concepts are, in practice within any control project, difficult to be explicitly used both for the analysis and design phases because of the lack of control-oriented tools that were integrated in agents design, and also because of the lack of type data abstraction (TDA) of CACSD systems. In this work Matlab/Simulink is used because it contains matrices TDA and also is a control-oriented programming language. It is specifically designed to help in the three phases of control engineering projects, both as a continuous-time simulation framework for testing the integrated simulated solution and also as a control system for micro-robots.
1.3. Multi-agents The next basic elements of the multi-agent frameworks can be distinguished: a common environment, agents, interaction between agents, and interactions between agents and the environment. In this work, the environment is the soccer field that is shared by another soccer team that continuously evolves in it, so that it must be considered as a dynamic environment: positions and velocities are continuously changing. The team is composed of three similar agents, whose behaviour can vary widely according to the instantaneous state of their internal parameters~ and depending on the value of two particular fuzzy variables: distance to the ball (DPB) and distance from the goal (DPG). It seems natural that only close the neighbourhood of every agent is of immediate relevance to that agent, so that the size of this neighbourhood is
ZLl. de la Rosa et al./Robotics and Autonomous Systems 2l (1997) 167-176
4, HOll
sp,~-,u--.,,,..,r,,.
] a,,.d,,~. ~ , , ~
aj ~ t ~ r o - r ~ , r
~.~3,~,5 s.~ a) RuI~I Ot' MIROSOT
b) A I n U ~ r m c h
i
:
¢) A Ffwlnework C'k~er to I A~U~-nmlcC~lml /
b
c,,~=.os, c~~ c , , . , , ~ .. ................).-if,;;,;;, e) MATLAB-SIMULn'qK
~
I
5.-~,,,-~o, I
I,~ P.. c ~ . l~ ~ " ~ '
quires every agent to have a model of a message recipient and to infer the impact of a message on the recipient based on that model. In this work, communication primitives consist of a subset of those used in AGENT0 [131, so that timely decision making is intended.
I /
f
~D,,a,
I
I
'. ~"~-"~
I
2. Agent-oriented
programming
] b) Co01~'lt['¢~ P*e.l..~falZ l ~..
I
F i g , 2. A n a l y s i s
I
l----v-~-s-~-.-~a-"............. l " (I)
169
o,,~o-suaJ~)
I - (2)
and design phases.
used to set a limit to the DPB variable range. By this way, decisions will be taken with degrees of certainty which can be zero out of the neighbourhood limits (see Fig. 2). Every agent includes perception and communication capabilities, so as decision capabilities. Since all the micro-robots of this project have the same technical specifications, various reactive models (behaviour roles, as for instance, defence, attack, and goal keeper) are programmed on them to avoid interfering each other, and to co-operate when they could mutually benefit. An important capacity of an agent in multi-agent environments is the ability to decide its own actions based on its own goals. However, when considering the relationships between individual goals and community goals two issues arise: how the satisfaction of individual goals affects the satisfaction of community goals, and how the satisfaction of community goals leads to the satisfaction of the individual goals. In this work, the satisfaction of community goals is guaranteed because final decisions are criticised by a fusion procedure, which can be understood as a private action of a new agent so-called coach-agent. The role of this agent is to review the set of individual decisions and make some changes only when needed (see Section 2.2.3). When operating in multi-agent environments, intelligent agents must generally co-ordinate their actions; thus they must communicate the proper knowledge and information. To develop techniques for deciding what to communicate is problematic [8] because it re-
AOP can be intuitively viewed as a specialisation of the object-oriented programming (OOP). Whereas, on the one hand, the OOP proposes to conceive software applications made of information modules or structures (so-called objects) that are able to exchange information between each other and that have individual ways of handling incoming messages by means of the so-called methods. On the other hand, AOP specialises the OOP framework by extending the state of the objects that now will be considered as agents with mental states (which consist of components such as beliefs, capabilities, and decisions). Moreover, the AOP contains further advanced and specialised methods of message passing between agents, as for example: to inform, request, offer, accept, reject, compete, and assist each other. The declarative knowledge (rules of reasoning) of agents has to be implemented using Matlab/Simulink which does not include object-oriented programming capabilities. Therefore, differences between both paradigms (agents and simulation/control) in the analysis phase are solved by splitting objects and developing reasoning in a sequential order. Thus, reasoning is developed in three steps: (1) Each agent decides its own reactive action depending on its position on the ground and the relative situation of the ball. Then, they inform to any agent this decision. (2) Each agent decides its cognitive action. Agents get new information and take new decisions (cooperative - cognitive - ones) that have higher degree of certainty than the reactive ones [5]. Then, they inform to coach-agent. (3) Individual decisions of micro-robots are criticised by coach-agent and converted into actions by selecting from proposals of the soccer agents. A simulation programme has been designed to test different reactive models, then any real micro-robots
170
J.LI. de la Rosa et al./Robotics and Autonomous Systems 21 (1997) 167-176
team runs with different co-ordination patterns by changing it own reactive models.
Table I Set of actions Private
Communicative
SHOOT GET FORW BACK
INFORM REQUEST
2.1. The AGENTO
The emergence of a number of prototypical agent languages [4,5,14] is one sign that agent technology is more widely used, and that many more agent-based applications are likely to be developed in the near future. One such language is Shoham's AGENT0 system, which is the inspiration of this work. In this language, an agent is specified in terms of sets of: (1) Agent's capabilities: actions the agent is able to perform. (2) Initial beliefs and commitments: logical statements about the world, that the agent believes to be true or false. (3) Commitment rules consist of antecedent conditions that are matched against incoming messages and the agent's internal states. The key component, which determines how the agent acts, is the commitment rule set. Each commitment rule contains a message condition, a mental condition, and an action. To determine whether any rule fires, the message condition is matched against the messages the agent has received; the mental condition is matched against the beliefs of the agent. If the rule fires, then the agent becomes committed to the action. Actions may be private, corresponding to an internally executed subroutine, or communicative, i.e., sending messages. Soccer players must be able to receive information about the environment (ball and players of the other team) and from the other players of the same team. The decision of every soccer player uses all this information as the basis of the reflexive reasoning and particular actions must be taken. Six actions, four private and two communicative, are defined for soccer players (see Table 1). • Private actions: shoot the ball (SHOOT), get the ball (GET), move forward (FORW), and go backwards (BACK). • Communicative actions: send a decision to a specified soccer player (INFORM), and request an action to another soccer player (REQUEST). When dealing with reactive systems, the controller is forced to send control signals to the system in every sample time and, in the soccer team this implies
that the communicative actions cannot take a long time. Communication between agents could present sequences of commitments and timing problems appear when this communication takes longer than usual. Paying attention to that case, the number of commitments is constrained in every grain of time to be consistent with reaI time features. Sometimes conflictive or unsolved situations appear, so actions are criticised [5] or reviewed by another co-operative agent, which has a globai view of situations and suggests solutions to conflicts. Communicative actions involve changes on the certainty degrees of private actions so that fusion procedures are used to calculate the final actions in a more reliable and confident way. Initial beliefs and agent capabilities can change according to the variation of the team strategy. In this work, databases are not used to store the whole history of the system, but destructive assignment (variables) are used to store only the 'recent' history, 2.2. Step-by-step reasoning
As mentioned above (see Section 2), the reasoning procedure to control this soccer team has been accomplished by using three steps. Next section describes how these steps work upon AGENT0 syntax, as well as some examples. 2.2.1. Reactive decisions
In the first step of the reasoning procedure, every soccer player decides a private action instinctively. This decision depends on local environment configuration (BELIEFS) defined by two parameters: distance player-ball (DPB), and distance player-goal (DPG). Joint to this private action, the decision also has a degree of certainty (see Fig. 3) to increase knowledge about the others.
171
J.Ll. de la Rosa et al./Robotics and Autonomous Systems 21 (1997) 167-176
3•
(3a)
ZONE 1
cation channel. Thus, reactive behaviour is developed and create rough intentions.
ZONE2
l
,~ I
-..\
//
. ~ D F B " ""..
i
.
~tzo~.,:0.8
,/"
"'~
" -.
'~,
DPB
DPB : point '2' • k '\
C..._:: .
( ,
/ ZONE I /
', "
....
"
~
ZONE2 Neigbourhood
.......................
(3b)
Attack Goal L._.__.2
NEAR f
/
limitation
ZONE3
FAR
',~ ;,
ttF^~0.7
DPG o
.
~_
DPG=point 'M'
\
DefenseGoal
Field limitation
Fig. 3. Reactive reasoning of the agents upon fuzzy sets: (a) with DPB variable; (b) with both DPB and DPG variables.
As an example, Fig. 3(a) shows a top view of the micro-robot with the ball in three possible situations. Decisions will be: SHOOT when at '1', GET when at '2', and FORW or BACK when at '3' depending on DPG value. By this way, the first reasoning procedure would be expressed like rule 1: Rule 1. B E L ( A g e n t X , DPB, ZONE2) ~INFORM ( to-any-agent, AgentX, SHOOT, 0.8)
2.2.2. Co-operative decisions This step implements the cognitive reasoning. It begins with a REQUEST (for communication) action, so that every agent can know the set (reactive_action, certainty, ID_player) of all other playmates. Therefore, when two playmates take conflictive decisions, certainty degrees are taken into account and one of them changes the former reactive decisions. Finally, agents communicate the decision to the coach-agent by an INFORM action. Fig. 4 shows a situation where both Agentl and Agent2 decide to GET the ball. After REQUEST themselves, Agentl will change to another action because its DPB parameter brings less certainty than those obtained by Agent2. This reasoning procedure could be expressed as I N F O R M (A g e n t 2, A g e n t i, B E L (Agent2, SHOOT, 0.7 } ) A B E L (Agentl, SHOOT, 0.2 ) B E L (Agentl, SHOOT, f (0.7,0.2) ) , where
f
(cl, c 2 ) = [ 0 ' I c2,
cl > c 2 , otherwise.
In this situation, since f ( c l , c2) -- 0, Agentl will change to FORW or BACK action using rules like 'rule 2'. This kind of communication is an exchange of information, and the knowledge of the environment increases. In that case, formed reactive decisions become co-operative ones.
Similarly, when at '3' in point 'M' (see Fig. 3(b)), ~asoning would b e a s m l e 2 :
Example: Both believe 'GET' /
but DPBI>DPB2
KEAC~ ZV]I 8TRlP
Rule 2.
Agent 1
~C(~t~,aZT.
B E L (AgentX, DPB, ZONE3 ) A B E L (AgentX, DPG, FAR) I N F O R M (to_ a n y - a g e n t , A g e n t X , FORW, c e r t a i n t y ), where 'certainty' is the final value obtained by
fuzzy inference calculation. After this first decision is taken, agents communicate it (INFORM) to the others through the communi-
.,,
0.2)
. " ~ .... ~--....
..../,-... .:>,~'.
BZL (AGent1,
J'C,;IO~, oz IIEL(AGmat:I,IIACK,
\,
7' ) ? )
Agent2 ItZL (A~len~.2, GlgT, 0 , 7 ) X• \
" ....
'
-
/
BEL (Agez~C2, GIRT, 0 . "7 )
//
"'~. ................ s "~
Fig. 4. Example of co-operative decision.
172
J.LI. de la Rosa et al./Robotics and Autonomous Systems 21 (1997) 167-176 •
/'" ...... ""
Set-Point position ( l, l )
,.";.---2k£:~-5-" Agent3 Agent2
;
'.
k
<,,.
,
l.,
I
""........"" Agentl
Fig. 5. Who GET the ball.
2.2.3. Fusion procedure Since this reactive team might be myopic in their approach, global information is useful and it is in this capacity that the coach-agent can interact with this multi-agent control team [2]. After co-operative set of decisions are received by the coach-agent, they are reviewed by a fusion procedure, which could be understood as a private action. Taking into account that this agent has a global overview of the environment, this difference in scope allows to solve remaining unsolved problems. This agent also contains some strategy parameters and can exert its influence based on a strategy of the match. Fig. 5 shows a usual indecision. If none of the agents decides the 'GET' action but the strategy game needs to get it, the coach-agent will select the Agent2, which has the best position to get the ball. Only the coachagent can evaluate this kind of situations because of its global scope, that is to say, it only knows where is the ball and where are the playmates. By this way, the role of the coach-agent is to change some co-operative decisions only if needed, and then REQUEST them. By this way, the private action of Agent2 could be as follows: REQUEST(Coach,Agent2, =~ D O ( A g e n t 2 ,
DO (GET))
SHOOT)
2.3. Local motion Decisions are high level descriptions of the control signals to be executed by the micro-robots, so that, individual set-point positions must be calculated. In this work, the close neighbourhood of each micro-robot is described by matrices. Therefore, relative set-point positions can be described as individual matrix elements, and are calculated by means of qualitative control [ 1,6]. By this way, the close neighbourhood space is split and facilitates obstacle avoidance algorithms. Fig. 6 shows typical simple path planning with an ob-
Fig. 6. Obstacle commitments.
avoidance
delays
the
achievement
of
stacle detected at 'A' point that delays the achievement of committments. This delay highlights the importance of persistence in intention to achieve goals. However, the real implementation contains this planning in HOST due to the limited amount of memory in the micro-robots microcontroller.
3. Multi-agent paradigm using Matlab/Simulink Step-by-step reasoning brings the opportunity to use modularised languages as Matlab/Simulink is. In this case, graphical tool provided by SIMULINK are useful because every step of reasoning can be programmed as a sequence of blocks diagram. As an example, Fig. 7 shows the blocks diagram corresponding to the reactive decision block. Then, previsouly defined DPG and DPB parameters are calculated so that a MATLAB-function block executes a fuzzy reactive action (i.e., action and its degree of certainty). With this graphical language, communication is performed using input and output arrows. Therefore, communicative actions as INFORM and REQUEST are easily implemented (see Fig. 8). An static table is
XY PLAYER
i
-~ XY REDGOAL
Xp-Xi
O~
~
I .I-"-]
~ Mux Xg-X~
MA~n-AB '''
[]
REACTIVE v"ector--"(2~, DECISION
DPB
Fig. 7. First decision (PLY_X block in Fig. 8).
J.LI. de la Rosa et al./Robotics and Autonomous Systems 21 (1997) 167-176
E~XY Ball
173
STEP-BY-STEPREASONING i
--
--
AGENT
SIMULATION ANIMATION BLOCK
FM EMITTER BLOCK
~_~ Servo3
Fig. 8. Blocks diagram for step-by-step reasoning.
placed between the first and the second steps of reasoning, and message passing is allowed (block 'TABLE'). Since this structure of communication could be understood as a basic blackboard-based scheme, the aim is to provide a high level communication channel widely used in agent paradigm. The implementation of all reasoning procedures is fully depicted in Fig. 8, where PLY-X blocks are the reactive decision blocks of players 1, 2, and 3. The way that players send (INFORM) their decisions to the 'TABLE' block is implemented by using a MUX block, so that the decision set comes into the 'TABLE' block as a vector. In the next step, players will receive (REQUEST) that set of decisions from the 'TABLE' block as a vector, too. Using the same vector representation, co-operative decisions come into the 'COACH-AGENT' block to be computed by a fusion procedure. This procedure consists of a sequential program that takes into account different strategy parameters: usual positions for every player, necessity to get the ball, and so on. Next, the coach-agent performs the COMMITMENT or the INFORM to-any-agent. Finally, a new set of decisions goes from the 'COACH-AGENT' block to every player and, locally, each one converts its final decision into a set-point position. This set-point is then executed by low level micro-robot motion architectures. When the simulation facility is used, an animation block is used to simulate the micro-robot movements. In this case, a cinematic model of the micro-robot is
introduced into another block and current positions can be calculated. But in real essays, current position values are provided by a vision system, and set-point positions will be sent to each micro-robot by the HOST computer ('FM EMITTER' block). Fig. 1 shows a general scheme of the whole system, and the connections between its elements. So far, it is important to notice that communication between robots themselves is not available; thus co-operative decisions are taken in the HOST computer. Therefore, since communicative actions related to reasoning process are designed by connecting blocks, an INFORM action from coach-agent to any-agent is implemented by using radio-communication.
4. Real system implementation Preceding sections described how private actions are decided using Matlab/Simulink framework. This section describes hardware, as well as special boards used for vision and communication. Paying attention to Fig. 8, the system description also includes micro-robot technical specifications split in PLY_#
and PlyCoop#. A main reason to use Matlab/Simulink software yields on the fact that one of its facilities, the RealTime Workshop, brings the opportunity to generate C-code ready to be compiled and executed even with real-time features. In this work, WATCOM C / C + +
174
J.L1. de la Rosa et al./Robotics and Autonomous Systems 21 (1997) 167-176
............................i M ~ ' ~ ; ~ 3 i ~ ...............................i i ~ ...................i ~ ~ s ~ ................ i ........................................................................................................................i i i............................................................ i l-imtw~-c Implementation 7. Micro-robot
!i
a) FM (Frequency Modulation) Half Duplex. RS232 b) RT Colou~ Segmemation. RTC + PIP MATROX c) INTEL 8751. Executing orders from HOST d) Mini-Motors. PWM. Power Source. Infrared Sensor
(1) ~
I
I (2) ~
!
,. co,,,..,,, ,,oq
¢) AOP/OOP ~m~p~edto
^ni..,o.
~. ~
Vector orientation
i Position : gravity center Imitation of the i ~ ~ork
a) Isolation of the '~
I
[
Reasoning/Control core from the simuime,,d
Fig. 10. Micro-robot top view colours.
environment
b) C~m~ation of Real Time C Co~k~fimm M A T L A B SIMULINK
to the vision system to obtain position and orientation of our team micro-robots.
c) Possible
use of MatlebSimulink interface for direct con~ol. d) Sl~eazial]singknowledge between H O S T and Micro-robot
l
Fig. 9. Implementation and integration phases. compiler I is used and constrained to obtain MS-DOS executable files. Also, another important feature is that specific drivers can be programmed to connect Matlab/Simulink to a hardware computer, like frame grabbing board, RS-232C serial communication port, I/O boards, and so on (see Fig. 9).
4.2. Communication system
Co-operative commands which are finally decided by the coach-agent are broadcasted to the robots by a FM radio-communication with a specially devised protocol. Because the FM emitter 4 can be connected to the serial communication port of the computer, co-operative commands can be sent directly from Simulink model. Notice that communication problems can affect the behaviour, also the performance of the robots because the lack of this communication induces reactive actions instead of co-operative ones.
4.1, Vision system
4.3. Micro-robot structure
The reasoning procedure needs to know the set of positions, speed and orientation of the robots, so as the ball speed and ball position. Since speed can be calculated using preceding position values, current values are provided by a vision system. In this work, the vision system is composed of a CCD camera, a specialpurpose real-time image co-processor (RTC), 2 and a PIP-1024B frame grabbing board. 3 The CCD video camera send images to the RTC which execute colour segmentation operations in realtime. After these operations, the frame grabbing board gets the image and digitises it to calculate gravity centres of segmented objects. As can be seen in Fig. 10, micro-robots have two colours to give important help
According to MIROSOT rules, robots have strong constraints on their size so that communication features and autonomous motion capabilities must be implemented by using small devices. Micro-robots are composed of one FM receiver, 5 a microprocessor 8751, a power unit, two micro-motors, 6 and batteries for motion, communication and control unit supply voltages (Fig. 1 l). When the communication unit receives information, the FM receiver converts it to RS-232C protocol and sent it to the on-board microprocessor. After programme calculations, control signals are generated and finally executed by PWM modulation.
1Watcom C/C++ (ver. 10.5), Tenberry Software, Inc., 1994. 2Real-time video co-processor (RTC) was developed by Vision and Group of Electronics, Informatics and Automatics Department of University of Girona. 3 PIP-1024B digitizer plug-in board is a comercial product of MATROX.
4Radiofrequency transmitter module, MOD: TX-433-SAW, 433,92 MHz, Totem Line. AUR-EL S.p.A. 5 Super-Heterodyne receiver with monolithic filter oscillator, MOD: STD 433 SIL. 433,92 MHz, Totem Line. AUR-EL S,p.A. 6 Maxon DC Motor, MOD: 2017.938-22.162-000. 12 V, 1.5 W 148000rpm., and Maxon gear, MOD:P 2916.804-0123.0-000, 122,6:1.
J.L1. de la Rosa et al./Robotics and Autonomous Systems 21 (1997) I67-176
Fig. 11. The implemented soccer micro-robots.
The main program enables two interrruption lines: a RS-232C port line, and an external one. With the first line, new commands from the host can be received so that new relative set-point positions are accepted. On the other hand, the external interrruption line is connected to an infrared sensor in order to avoid collisions or finely detect the ball.
5. Conclusions
This project is quite stimulating in several terms. With respect to the technical features of the proposed solution, the Matlab/Simulink integrated vision system is able to calculate position and orientation of the ball and micro-robots taking approximately 750ms. long. The micro-robots are not as autonomous as desired to apply agents' approaches on stand-alone. The system was applied on-line without the immediate need of generating C code. With respect to the research on this project, the first idea from the analysis phase is that agents are natural way of thinking about distributed, specialised autonomous systems. To deal with this aim two interesting results have appeared. The former, the Shoham selected AOP [13], is quite general and basically enough for our system; and though implementation of cooperation is done in the HOST, the conceptual features are cooperative. Let us explain further on this: the analysis of this project is a subset of SHOHAM AGENT0, and anyway, the solution is proper. In the design phase the subset of SHOHAM was equally enough. In the
175
implementation phase, most of the responsibility for the co-operation is supported by the HOST system mainly due to the half-duplex nature of the communication device finally implemented in our robots, and also by the limited capabilities of the chosen INTEL 8751 micro-processor. Furthermore, since these micro-robots are not fully sensorised, this is not, so far, a loss of generality of the final solution. The latter, the necessity of implementing AOP within a CACSD framework closer to control engineers, is justified for facilitating the analysis and design phases because, first, of the continuous time dynamic simulation facilities of the selected CACSD system, Matlab/Simulink, and second, because the amount of control oriented toolboxes that are developed from current research to this environment suggest a brilliant platform for the application of AOP in automatic control and automation. However, implementing the AOP is not easy because this is not an OOP framework. One conclusion is that Matlab/Simulink could extremely help the analysis and design phases if this system was object-oriented. Nowadays, some works try to deal with this lack [9,10]. On the other hand, in the implementation phase this tool is now quite powerful to generate C software code (both non-real and real-time operation) which is extremely interesting in obtaining a final integrated automatic control-oriented system [12]. Therefore, this could be a first step to a general application of agents in the world of automatic control and robotics by using tools that are common in the former area. The drawback is that object-oriented paradigms are not extended enough and agents have problems to be applied in practical implementations. This paper shows a possibly first success in practical application/implementation of agents in automatic control.
6. Future work
Further work on Matlab/Simulink must improve support object-oriented tools so that research community could go on to research supervisory control architectures, methodologies, and practical applications. A near future version Matlab/Simulink is announced to incorporate software structures. Structures will facilitate the application of these ideas to embed objects
176
J.LI. de la Rosa et al./Robotics and Autonomous Systems 21 (1997) 167-176
better. H o w e v e r , these structures are e x p e c t e d to bec o m e real objects in the next versions as required for AOP.
Acknowledgements This w o r k has b e e n partially f u n d e d by the TAP961114-C03-03 project Plataformas Integradas de C A D de Supervisi6n y Metodologfas o f C I C Y T p r o g r a m f r o m the Spanish g o v e r n m e n t .
References [1] N. Agell, J.C. Aguado and N. Piera, A qualitative mobil tracking, Current Trends in Qualitative Reasoning and Applications, Monograph CIMNE No 33 (1995) 102-107. [2] R.C. Arkin and K.S. Ali, Integration of reactive and telerobotic control in multi-agent robotic systems, Proc. 3th Int. Conf, on Simulation of Adaptive Behaviour. [3] T. Balch, G. Boone, T. Collins, H. Forbes, D. Mackenzie and J.C. Santamaria, Io, Ganymede and Callisto: A multiagent robot trash-collencting team, AI Magazine (16) (2) (1995) 39-51. [4] COORDINATION: Linguistic support for multiple cooperating agents, ESPRIT Project No 9102 in the Basic Research Action, Work Area: Basic Aspects of Multiple Computing Systems, Dates: 1.2.94-31.1.96. [5] J.Li. de la Rosa, J. Aguilar and I. Serra, Heuristics for Cooperation of Expert Systems. Application to Process Control, Ed. PIAR, 1994, ISBN 84-605-0275-9. [6] J.LI. de la Rosa, A. Oller et al., A comparison of fuzzy and qualitative control techniques, IEEE Int. Conf. on Control Applications~Int. Symp. on Intelligent Control~Int. Symp. on Computer-Aided Control System Design (ISBN 0-78032978-3) Dearborn, USA (1996) t39-144. [7] T. Finin et al,, DRAFT specification of the KQML (The knowledge query & manipulation language), unpublished draft, 1993. [8] P.J. Gmytrasiewicz, E.H. Durfee and D.K. Wehe, The utility of communication in coordinating intelligent agents, Proc. 9th National Conf. on Artificial Intelligence ( 1991 ). [9] J. Melendez, J.L1. de la Rosa and J. Aguilar, Embedding objects into Matlab/Simulink for process supervision, IEEE Control Systems Magazine (1997), to appear. [10] J. Melendez, J.LI. de la Rosa, J. Colomer, J. Vehi and C. Pous, Incrustaci6n de objetos en Simulink. Integraci6n de herramientas de ayuda al disefio de estructuras de supervisi6n, in: The H Congreso de usuarios de MATLAB, Madrid, Spain, September 1996. [11] Micro-Robot Soccer team Tournament, 9-12 November 1996, KAIST, Korea.
[12] C. Pous, A. Oiler, J. Vehf, J.LI. de la Rosa, Using real-time workshop in teaching control design techniques, Volume on Real Time Systems Education (IEEE Computer Society Press, Silver Spring, MD, 1996, ISBN 0-8186-7649-3) 153-158, August 1996. [ 13] Y. Shoham, Agent-oriented programming, Artificial Intelligence 60 (I 993) 51-92. [I4] M. Wooldridge and N.R. Jennings, Intelligent agents: Theory and practice, Knowledge Engineering Review, (1995), submitted.
Josep Lluis de la Rosa i Esteva obtained a degree in Sciences at the Autonomous University of Barcelona, Bellaterra (Spain) in 1989. In 1989 he joined Siemens-Nixdorf SoIL ware Development Center in Barcelona. In 1991 joined the Laboratoire d'Architecture et d'Analyse des Systbmes of the Centre National de la Rechercher Scientifique (LAAS-CNRS) in Toulouse (France), where he got the Ph.D. degree in Sciences by the Autonomous University of Barcelona in 1993. From 1994 he is in the University of Girona, adscribed to the Institute of Informatics and Application (IliA). His present research interests include artificial intelligence applied to supervisory systems and process control, and multi-agents systems. WWW: http://eia.udg.es/'-~peplluis. Albert Oiler i Pujol obtained a degree in Sciences at the Autonomous University of Barcelona, Bellaterra (Spain) in 1990. From 1994 he is in the University of Girona, adscribed to the Institute of Informatics and Application (IliA). His present research interests include artificial intelligence applied to supervisory systems and process control, and multi-agents systems. WWW: http://eia.udg.es/"~oller.
Josep PuyoI-Gruart obtained a degree in Electronic Engineering at the Polytechnic University of Catalonia, Barcelona (Spain) in 1985. From 1987 he has worked for the Spanish Council for Scientific Research (CSIC) at the Artificial Intelligence Research Institute (IliA). He got the Ph.D. degree in Computer Science by the Autonomous University of Barcelona in 1994. His present research interests include knowledge-based systems and multi-agents systems. WWW: http://www.iiia.csic.es/~puyol.
Josep Vehl obtained a degree in Sciences at the Autonomous University of Barcelona, Bellaterra (Spain) in 1987. Since September 1987 he has worked at the Technical University of Catalonia. Since 1991 he has been working at the University of Girona, adscribed to the Institute of Informatics and Applications (IIiA). His present research interests include interval methods applied to robust control parametric design and interval simulation.
ELSEVIER
Soccer team based on agent-oriented programming J.L1. d e l a R o s a a,., A. O i l e r a, j. V e h i a, j. P u y o l b a Intelligent Systems and Control Engineering Group, Institute oflnformatics and Applications (HiA). University of Girona, Catalonia, Spain b Artificial Intelligence Research Institute (IliA), CSIC, Catalonia, Spain
Abstract
In this paper the analysis, design and implementation of a soccer team of micro-robots is explained. Besides the technical difficulties to develop these micro-robots, this paper also shows how to develop a multi-agent co-operative system by means of Matlab/Simulink (MATLAB and SIMULINK are Trade Marks of Math Works. Windows 95 is a Trade Mark of Microsoft Corporation.) a widely known computer aided control system design framework. Agent-oriented paradigms (AOPs) formalise interactions between multiple agents in terms of changing their mental states by communication between agents. Their practical implementations are usually conceived by means of object-oriented paradigms. Nevertheless, the implementation of agentoriented paradigms in Matlab/Simulink is not straightforward. Thus, the obtained real implementation is an integrated system that includes several programming paradigms so as hardware platforms. Finally, the proposal of the integrated framework for the micro-robots soccer team is shown. Keywords: Agents; CACSD; Soccer
1. Introduction
1.1. Mirosot
Multi-agent based mobile robotics require new examples from application and call for new control schemes. By this way, Micro-Robot World Cup Soccer Tournament (MIROSOT) [11] meeting is a forum where rules and constraints about the way that multiple micro-robots could work together in a game is a chance to apply techniques focused on agents. This game is the micro-robotic soccer where three micro-robots have to play against another team of the same characteristics and try to score as much as possible.
According to MIROSOT specifications, soccer micro-robots are limited to the size 7.5 × 7.5 x 7.5 cm, and must run autonomously on a 130 × 90 cm sized ground. MIROSOT rules have strong analogy with real soccer rules although adapted to current robotics technology. Teams can use a centralised vision system to provide robots with situations in terms of their position and orientation so as with the ball position. The vision-based facility was widely used by participants to the MIROSOT tournament. The vision-based measures are used by a computer host, here called HOST, to calculate individual robotic movements that are send (broadcasted by an FM emitter) to the robots to be executed. This is depicted in Fig. I. Thus, since robots have autonomous behaviour but the whole team
* Corresponding author. 0921-8890/97/$17.00 t997 Published by Elsevier Science B.V. PH S0921-8890(97)00024-9
J.Ll. de la Rosa et al./Robotics and Autonomous 5),stems 21 (1997) 167-176
t68
...',./:- !~ "~:..":. . . . . . . . . . . . . . .
i ..
VIDEO CAMERA CCD
........ "-..,.
R O B O T C...,OLORS ( T O P VIEW )
7.5 crn, I .: .~ : .,: ,i:
~ TEAMCOLOR (frontpart) .*---- PLAYER COLOR (bact~part)
Fig. I. The overall system. must actuate as a group, they must co-operate. This means that robots must communicate each other, and the proposal of this work is to deal with a more general problem, that is to do reasoning on communication as agents do. Therefore, agent-oriented paradigm (AOP) frameworks [13] are used in this work.
1.2. Agent-oriented paradigms and control Nowadays, a commonly proposed solution to the behaviour control in a community of robots is the application of agents theory [3]. The control of the global behaviour of a soccer team problem gives the opportunity to apply agents theory due to the distributed architecture of the mobile robots and the problem itself that implies coordination, competition, and cooperation by means of communication among each soccer player and perception. Therefore, the communication of information and knowledge is the most important action that might be taken by each agent, in this case, each soccer player. On the other hand, control engineering projects are developed as usual in the following steps: analysis, design, and implementation. There are modern computer-based tools that handle the control engineering practice, named computer aided control systems design (CACSD) environments, The robotics applications could be more easily developed by using such environments because of the availability of linear and non-linear systems analysis and design tools that could be useful to control robotics movements.
These two approaches are clearly different: the AOP provides us the conceptual model to describe, in a declarative style, the behaviour of the team by means of high level constructs and the current CACSD are not good at representing declarative knowledge but procedural. When dealing with complex systems further declaration is fundamental. Therefore, at the moment to solve complex systems control then the distribution of knowledge is essential in addition to automatic control, and the case of the soccer micro-robotics game is specially adequate. Therefore in this paper the conjunction of AOP and CACSD systems is proposed, because the AOP introduces declarative knowledge in terms of communication operators for dealing with decisions, capabilities and commitments of agents that are necessary in this project. But these concepts are, in practice within any control project, difficult to be explicitly used both for the analysis and design phases because of the lack of control-oriented tools that were integrated in agents design, and also because of the lack of type data abstraction (TDA) of CACSD systems. In this work Matlab/Simulink is used because it contains matrices TDA and also is a control-oriented programming language. It is specifically designed to help in the three phases of control engineering projects, both as a continuous-time simulation framework for testing the integrated simulated solution and also as a control system for micro-robots.
1.3. Multi-agents The next basic elements of the multi-agent frameworks can be distinguished: a common environment, agents, interaction between agents, and interactions between agents and the environment. In this work, the environment is the soccer field that is shared by another soccer team that continuously evolves in it, so that it must be considered as a dynamic environment: positions and velocities are continuously changing. The team is composed of three similar agents, whose behaviour can vary widely according to the instantaneous state of their internal parameters~ and depending on the value of two particular fuzzy variables: distance to the ball (DPB) and distance from the goal (DPG). It seems natural that only close the neighbourhood of every agent is of immediate relevance to that agent, so that the size of this neighbourhood is
ZLl. de la Rosa et al./Robotics and Autonomous Systems 2l (1997) 167-176
4, HOll
sp,~-,u--.,,,..,r,,.
] a,,.d,,~. ~ , , ~
aj ~ t ~ r o - r ~ , r
~.~3,~,5 s.~ a) RuI~I Ot' MIROSOT
b) A I n U ~ r m c h
i
:
¢) A Ffwlnework C'k~er to I A~U~-nmlcC~lml /
b
c,,~=.os, c~~ c , , . , , ~ .. ................).-if,;;,;;, e) MATLAB-SIMULn'qK
~
I
5.-~,,,-~o, I
I,~ P.. c ~ . l~ ~ " ~ '
quires every agent to have a model of a message recipient and to infer the impact of a message on the recipient based on that model. In this work, communication primitives consist of a subset of those used in AGENT0 [131, so that timely decision making is intended.
I /
f
~D,,a,
I
I
'. ~"~-"~
I
2. Agent-oriented
programming
] b) Co01~'lt['¢~ P*e.l..~falZ l ~..
I
F i g , 2. A n a l y s i s
I
l----v-~-s-~-.-~a-"............. l " (I)
169
o,,~o-suaJ~)
I - (2)
and design phases.
used to set a limit to the DPB variable range. By this way, decisions will be taken with degrees of certainty which can be zero out of the neighbourhood limits (see Fig. 2). Every agent includes perception and communication capabilities, so as decision capabilities. Since all the micro-robots of this project have the same technical specifications, various reactive models (behaviour roles, as for instance, defence, attack, and goal keeper) are programmed on them to avoid interfering each other, and to co-operate when they could mutually benefit. An important capacity of an agent in multi-agent environments is the ability to decide its own actions based on its own goals. However, when considering the relationships between individual goals and community goals two issues arise: how the satisfaction of individual goals affects the satisfaction of community goals, and how the satisfaction of community goals leads to the satisfaction of the individual goals. In this work, the satisfaction of community goals is guaranteed because final decisions are criticised by a fusion procedure, which can be understood as a private action of a new agent so-called coach-agent. The role of this agent is to review the set of individual decisions and make some changes only when needed (see Section 2.2.3). When operating in multi-agent environments, intelligent agents must generally co-ordinate their actions; thus they must communicate the proper knowledge and information. To develop techniques for deciding what to communicate is problematic [8] because it re-
AOP can be intuitively viewed as a specialisation of the object-oriented programming (OOP). Whereas, on the one hand, the OOP proposes to conceive software applications made of information modules or structures (so-called objects) that are able to exchange information between each other and that have individual ways of handling incoming messages by means of the so-called methods. On the other hand, AOP specialises the OOP framework by extending the state of the objects that now will be considered as agents with mental states (which consist of components such as beliefs, capabilities, and decisions). Moreover, the AOP contains further advanced and specialised methods of message passing between agents, as for example: to inform, request, offer, accept, reject, compete, and assist each other. The declarative knowledge (rules of reasoning) of agents has to be implemented using Matlab/Simulink which does not include object-oriented programming capabilities. Therefore, differences between both paradigms (agents and simulation/control) in the analysis phase are solved by splitting objects and developing reasoning in a sequential order. Thus, reasoning is developed in three steps: (1) Each agent decides its own reactive action depending on its position on the ground and the relative situation of the ball. Then, they inform to any agent this decision. (2) Each agent decides its cognitive action. Agents get new information and take new decisions (cooperative - cognitive - ones) that have higher degree of certainty than the reactive ones [5]. Then, they inform to coach-agent. (3) Individual decisions of micro-robots are criticised by coach-agent and converted into actions by selecting from proposals of the soccer agents. A simulation programme has been designed to test different reactive models, then any real micro-robots
170
J.LI. de la Rosa et al./Robotics and Autonomous Systems 21 (1997) 167-176
team runs with different co-ordination patterns by changing it own reactive models.
Table I Set of actions Private
Communicative
SHOOT GET FORW BACK
INFORM REQUEST
2.1. The AGENTO
The emergence of a number of prototypical agent languages [4,5,14] is one sign that agent technology is more widely used, and that many more agent-based applications are likely to be developed in the near future. One such language is Shoham's AGENT0 system, which is the inspiration of this work. In this language, an agent is specified in terms of sets of: (1) Agent's capabilities: actions the agent is able to perform. (2) Initial beliefs and commitments: logical statements about the world, that the agent believes to be true or false. (3) Commitment rules consist of antecedent conditions that are matched against incoming messages and the agent's internal states. The key component, which determines how the agent acts, is the commitment rule set. Each commitment rule contains a message condition, a mental condition, and an action. To determine whether any rule fires, the message condition is matched against the messages the agent has received; the mental condition is matched against the beliefs of the agent. If the rule fires, then the agent becomes committed to the action. Actions may be private, corresponding to an internally executed subroutine, or communicative, i.e., sending messages. Soccer players must be able to receive information about the environment (ball and players of the other team) and from the other players of the same team. The decision of every soccer player uses all this information as the basis of the reflexive reasoning and particular actions must be taken. Six actions, four private and two communicative, are defined for soccer players (see Table 1). • Private actions: shoot the ball (SHOOT), get the ball (GET), move forward (FORW), and go backwards (BACK). • Communicative actions: send a decision to a specified soccer player (INFORM), and request an action to another soccer player (REQUEST). When dealing with reactive systems, the controller is forced to send control signals to the system in every sample time and, in the soccer team this implies
that the communicative actions cannot take a long time. Communication between agents could present sequences of commitments and timing problems appear when this communication takes longer than usual. Paying attention to that case, the number of commitments is constrained in every grain of time to be consistent with reaI time features. Sometimes conflictive or unsolved situations appear, so actions are criticised [5] or reviewed by another co-operative agent, which has a globai view of situations and suggests solutions to conflicts. Communicative actions involve changes on the certainty degrees of private actions so that fusion procedures are used to calculate the final actions in a more reliable and confident way. Initial beliefs and agent capabilities can change according to the variation of the team strategy. In this work, databases are not used to store the whole history of the system, but destructive assignment (variables) are used to store only the 'recent' history, 2.2. Step-by-step reasoning
As mentioned above (see Section 2), the reasoning procedure to control this soccer team has been accomplished by using three steps. Next section describes how these steps work upon AGENT0 syntax, as well as some examples. 2.2.1. Reactive decisions
In the first step of the reasoning procedure, every soccer player decides a private action instinctively. This decision depends on local environment configuration (BELIEFS) defined by two parameters: distance player-ball (DPB), and distance player-goal (DPG). Joint to this private action, the decision also has a degree of certainty (see Fig. 3) to increase knowledge about the others.
171
J.Ll. de la Rosa et al./Robotics and Autonomous Systems 21 (1997) 167-176
3•
(3a)
ZONE 1
cation channel. Thus, reactive behaviour is developed and create rough intentions.
ZONE2
l
,~ I
-..\
//
. ~ D F B " ""..
i
.
~tzo~.,:0.8
,/"
"'~
" -.
'~,
DPB
DPB : point '2' • k '\
C..._:: .
( ,
/ ZONE I /
', "
....
"
~
ZONE2 Neigbourhood
.......................
(3b)
Attack Goal L._.__.2
NEAR f
/
limitation
ZONE3
FAR
',~ ;,
ttF^~0.7
DPG o
.
~_
DPG=point 'M'
\
DefenseGoal
Field limitation
Fig. 3. Reactive reasoning of the agents upon fuzzy sets: (a) with DPB variable; (b) with both DPB and DPG variables.
As an example, Fig. 3(a) shows a top view of the micro-robot with the ball in three possible situations. Decisions will be: SHOOT when at '1', GET when at '2', and FORW or BACK when at '3' depending on DPG value. By this way, the first reasoning procedure would be expressed like rule 1: Rule 1. B E L ( A g e n t X , DPB, ZONE2) ~INFORM ( to-any-agent, AgentX, SHOOT, 0.8)
2.2.2. Co-operative decisions This step implements the cognitive reasoning. It begins with a REQUEST (for communication) action, so that every agent can know the set (reactive_action, certainty, ID_player) of all other playmates. Therefore, when two playmates take conflictive decisions, certainty degrees are taken into account and one of them changes the former reactive decisions. Finally, agents communicate the decision to the coach-agent by an INFORM action. Fig. 4 shows a situation where both Agentl and Agent2 decide to GET the ball. After REQUEST themselves, Agentl will change to another action because its DPB parameter brings less certainty than those obtained by Agent2. This reasoning procedure could be expressed as I N F O R M (A g e n t 2, A g e n t i, B E L (Agent2, SHOOT, 0.7 } ) A B E L (Agentl, SHOOT, 0.2 ) B E L (Agentl, SHOOT, f (0.7,0.2) ) , where
f
(cl, c 2 ) = [ 0 ' I c2,
cl > c 2 , otherwise.
In this situation, since f ( c l , c2) -- 0, Agentl will change to FORW or BACK action using rules like 'rule 2'. This kind of communication is an exchange of information, and the knowledge of the environment increases. In that case, formed reactive decisions become co-operative ones.
Similarly, when at '3' in point 'M' (see Fig. 3(b)), ~asoning would b e a s m l e 2 :
Example: Both believe 'GET' /
but DPBI>DPB2
KEAC~ ZV]I 8TRlP
Rule 2.
Agent 1
~C(~t~,aZT.
B E L (AgentX, DPB, ZONE3 ) A B E L (AgentX, DPG, FAR) I N F O R M (to_ a n y - a g e n t , A g e n t X , FORW, c e r t a i n t y ), where 'certainty' is the final value obtained by
fuzzy inference calculation. After this first decision is taken, agents communicate it (INFORM) to the others through the communi-
.,,
0.2)
. " ~ .... ~--....
..../,-... .:>,~'.
BZL (AGent1,
J'C,;IO~, oz IIEL(AGmat:I,IIACK,
\,
7' ) ? )
Agent2 ItZL (A~len~.2, GlgT, 0 , 7 ) X• \
" ....
'
-
/
BEL (Agez~C2, GIRT, 0 . "7 )
//
"'~. ................ s "~
Fig. 4. Example of co-operative decision.
172
J.LI. de la Rosa et al./Robotics and Autonomous Systems 21 (1997) 167-176 •
/'" ...... ""
Set-Point position ( l, l )
,.";.---2k£:~-5-" Agent3 Agent2
;
'.
k
<,,.
,
l.,
I
""........"" Agentl
Fig. 5. Who GET the ball.
2.2.3. Fusion procedure Since this reactive team might be myopic in their approach, global information is useful and it is in this capacity that the coach-agent can interact with this multi-agent control team [2]. After co-operative set of decisions are received by the coach-agent, they are reviewed by a fusion procedure, which could be understood as a private action. Taking into account that this agent has a global overview of the environment, this difference in scope allows to solve remaining unsolved problems. This agent also contains some strategy parameters and can exert its influence based on a strategy of the match. Fig. 5 shows a usual indecision. If none of the agents decides the 'GET' action but the strategy game needs to get it, the coach-agent will select the Agent2, which has the best position to get the ball. Only the coachagent can evaluate this kind of situations because of its global scope, that is to say, it only knows where is the ball and where are the playmates. By this way, the role of the coach-agent is to change some co-operative decisions only if needed, and then REQUEST them. By this way, the private action of Agent2 could be as follows: REQUEST(Coach,Agent2, =~ D O ( A g e n t 2 ,
DO (GET))
SHOOT)
2.3. Local motion Decisions are high level descriptions of the control signals to be executed by the micro-robots, so that, individual set-point positions must be calculated. In this work, the close neighbourhood of each micro-robot is described by matrices. Therefore, relative set-point positions can be described as individual matrix elements, and are calculated by means of qualitative control [ 1,6]. By this way, the close neighbourhood space is split and facilitates obstacle avoidance algorithms. Fig. 6 shows typical simple path planning with an ob-
Fig. 6. Obstacle commitments.
avoidance
delays
the
achievement
of
stacle detected at 'A' point that delays the achievement of committments. This delay highlights the importance of persistence in intention to achieve goals. However, the real implementation contains this planning in HOST due to the limited amount of memory in the micro-robots microcontroller.
3. Multi-agent paradigm using Matlab/Simulink Step-by-step reasoning brings the opportunity to use modularised languages as Matlab/Simulink is. In this case, graphical tool provided by SIMULINK are useful because every step of reasoning can be programmed as a sequence of blocks diagram. As an example, Fig. 7 shows the blocks diagram corresponding to the reactive decision block. Then, previsouly defined DPG and DPB parameters are calculated so that a MATLAB-function block executes a fuzzy reactive action (i.e., action and its degree of certainty). With this graphical language, communication is performed using input and output arrows. Therefore, communicative actions as INFORM and REQUEST are easily implemented (see Fig. 8). An static table is
XY PLAYER
i
-~ XY REDGOAL
Xp-Xi
O~
~
I .I-"-]
~ Mux Xg-X~
MA~n-AB '''
[]
REACTIVE v"ector--"(2~, DECISION
DPB
Fig. 7. First decision (PLY_X block in Fig. 8).
J.LI. de la Rosa et al./Robotics and Autonomous Systems 21 (1997) 167-176
E~XY Ball
173
STEP-BY-STEPREASONING i
--
--
AGENT
SIMULATION ANIMATION BLOCK
FM EMITTER BLOCK
~_~ Servo3
Fig. 8. Blocks diagram for step-by-step reasoning.
placed between the first and the second steps of reasoning, and message passing is allowed (block 'TABLE'). Since this structure of communication could be understood as a basic blackboard-based scheme, the aim is to provide a high level communication channel widely used in agent paradigm. The implementation of all reasoning procedures is fully depicted in Fig. 8, where PLY-X blocks are the reactive decision blocks of players 1, 2, and 3. The way that players send (INFORM) their decisions to the 'TABLE' block is implemented by using a MUX block, so that the decision set comes into the 'TABLE' block as a vector. In the next step, players will receive (REQUEST) that set of decisions from the 'TABLE' block as a vector, too. Using the same vector representation, co-operative decisions come into the 'COACH-AGENT' block to be computed by a fusion procedure. This procedure consists of a sequential program that takes into account different strategy parameters: usual positions for every player, necessity to get the ball, and so on. Next, the coach-agent performs the COMMITMENT or the INFORM to-any-agent. Finally, a new set of decisions goes from the 'COACH-AGENT' block to every player and, locally, each one converts its final decision into a set-point position. This set-point is then executed by low level micro-robot motion architectures. When the simulation facility is used, an animation block is used to simulate the micro-robot movements. In this case, a cinematic model of the micro-robot is
introduced into another block and current positions can be calculated. But in real essays, current position values are provided by a vision system, and set-point positions will be sent to each micro-robot by the HOST computer ('FM EMITTER' block). Fig. 1 shows a general scheme of the whole system, and the connections between its elements. So far, it is important to notice that communication between robots themselves is not available; thus co-operative decisions are taken in the HOST computer. Therefore, since communicative actions related to reasoning process are designed by connecting blocks, an INFORM action from coach-agent to any-agent is implemented by using radio-communication.
4. Real system implementation Preceding sections described how private actions are decided using Matlab/Simulink framework. This section describes hardware, as well as special boards used for vision and communication. Paying attention to Fig. 8, the system description also includes micro-robot technical specifications split in PLY_#
and PlyCoop#. A main reason to use Matlab/Simulink software yields on the fact that one of its facilities, the RealTime Workshop, brings the opportunity to generate C-code ready to be compiled and executed even with real-time features. In this work, WATCOM C / C + +
174
J.L1. de la Rosa et al./Robotics and Autonomous Systems 21 (1997) 167-176
............................i M ~ ' ~ ; ~ 3 i ~ ...............................i i ~ ...................i ~ ~ s ~ ................ i ........................................................................................................................i i i............................................................ i l-imtw~-c Implementation 7. Micro-robot
!i
a) FM (Frequency Modulation) Half Duplex. RS232 b) RT Colou~ Segmemation. RTC + PIP MATROX c) INTEL 8751. Executing orders from HOST d) Mini-Motors. PWM. Power Source. Infrared Sensor
(1) ~
I
I (2) ~
!
,. co,,,..,,, ,,oq
¢) AOP/OOP ~m~p~edto
^ni..,o.
~. ~
Vector orientation
i Position : gravity center Imitation of the i ~ ~ork
a) Isolation of the '~
I
[
Reasoning/Control core from the simuime,,d
Fig. 10. Micro-robot top view colours.
environment
b) C~m~ation of Real Time C Co~k~fimm M A T L A B SIMULINK
to the vision system to obtain position and orientation of our team micro-robots.
c) Possible
use of MatlebSimulink interface for direct con~ol. d) Sl~eazial]singknowledge between H O S T and Micro-robot
l
Fig. 9. Implementation and integration phases. compiler I is used and constrained to obtain MS-DOS executable files. Also, another important feature is that specific drivers can be programmed to connect Matlab/Simulink to a hardware computer, like frame grabbing board, RS-232C serial communication port, I/O boards, and so on (see Fig. 9).
4.2. Communication system
Co-operative commands which are finally decided by the coach-agent are broadcasted to the robots by a FM radio-communication with a specially devised protocol. Because the FM emitter 4 can be connected to the serial communication port of the computer, co-operative commands can be sent directly from Simulink model. Notice that communication problems can affect the behaviour, also the performance of the robots because the lack of this communication induces reactive actions instead of co-operative ones.
4.1, Vision system
4.3. Micro-robot structure
The reasoning procedure needs to know the set of positions, speed and orientation of the robots, so as the ball speed and ball position. Since speed can be calculated using preceding position values, current values are provided by a vision system. In this work, the vision system is composed of a CCD camera, a specialpurpose real-time image co-processor (RTC), 2 and a PIP-1024B frame grabbing board. 3 The CCD video camera send images to the RTC which execute colour segmentation operations in realtime. After these operations, the frame grabbing board gets the image and digitises it to calculate gravity centres of segmented objects. As can be seen in Fig. 10, micro-robots have two colours to give important help
According to MIROSOT rules, robots have strong constraints on their size so that communication features and autonomous motion capabilities must be implemented by using small devices. Micro-robots are composed of one FM receiver, 5 a microprocessor 8751, a power unit, two micro-motors, 6 and batteries for motion, communication and control unit supply voltages (Fig. 1 l). When the communication unit receives information, the FM receiver converts it to RS-232C protocol and sent it to the on-board microprocessor. After programme calculations, control signals are generated and finally executed by PWM modulation.
1Watcom C/C++ (ver. 10.5), Tenberry Software, Inc., 1994. 2Real-time video co-processor (RTC) was developed by Vision and Group of Electronics, Informatics and Automatics Department of University of Girona. 3 PIP-1024B digitizer plug-in board is a comercial product of MATROX.
4Radiofrequency transmitter module, MOD: TX-433-SAW, 433,92 MHz, Totem Line. AUR-EL S.p.A. 5 Super-Heterodyne receiver with monolithic filter oscillator, MOD: STD 433 SIL. 433,92 MHz, Totem Line. AUR-EL S,p.A. 6 Maxon DC Motor, MOD: 2017.938-22.162-000. 12 V, 1.5 W 148000rpm., and Maxon gear, MOD:P 2916.804-0123.0-000, 122,6:1.
J.L1. de la Rosa et al./Robotics and Autonomous Systems 21 (1997) I67-176
Fig. 11. The implemented soccer micro-robots.
The main program enables two interrruption lines: a RS-232C port line, and an external one. With the first line, new commands from the host can be received so that new relative set-point positions are accepted. On the other hand, the external interrruption line is connected to an infrared sensor in order to avoid collisions or finely detect the ball.
5. Conclusions
This project is quite stimulating in several terms. With respect to the technical features of the proposed solution, the Matlab/Simulink integrated vision system is able to calculate position and orientation of the ball and micro-robots taking approximately 750ms. long. The micro-robots are not as autonomous as desired to apply agents' approaches on stand-alone. The system was applied on-line without the immediate need of generating C code. With respect to the research on this project, the first idea from the analysis phase is that agents are natural way of thinking about distributed, specialised autonomous systems. To deal with this aim two interesting results have appeared. The former, the Shoham selected AOP [13], is quite general and basically enough for our system; and though implementation of cooperation is done in the HOST, the conceptual features are cooperative. Let us explain further on this: the analysis of this project is a subset of SHOHAM AGENT0, and anyway, the solution is proper. In the design phase the subset of SHOHAM was equally enough. In the
175
implementation phase, most of the responsibility for the co-operation is supported by the HOST system mainly due to the half-duplex nature of the communication device finally implemented in our robots, and also by the limited capabilities of the chosen INTEL 8751 micro-processor. Furthermore, since these micro-robots are not fully sensorised, this is not, so far, a loss of generality of the final solution. The latter, the necessity of implementing AOP within a CACSD framework closer to control engineers, is justified for facilitating the analysis and design phases because, first, of the continuous time dynamic simulation facilities of the selected CACSD system, Matlab/Simulink, and second, because the amount of control oriented toolboxes that are developed from current research to this environment suggest a brilliant platform for the application of AOP in automatic control and automation. However, implementing the AOP is not easy because this is not an OOP framework. One conclusion is that Matlab/Simulink could extremely help the analysis and design phases if this system was object-oriented. Nowadays, some works try to deal with this lack [9,10]. On the other hand, in the implementation phase this tool is now quite powerful to generate C software code (both non-real and real-time operation) which is extremely interesting in obtaining a final integrated automatic control-oriented system [12]. Therefore, this could be a first step to a general application of agents in the world of automatic control and robotics by using tools that are common in the former area. The drawback is that object-oriented paradigms are not extended enough and agents have problems to be applied in practical implementations. This paper shows a possibly first success in practical application/implementation of agents in automatic control.
6. Future work
Further work on Matlab/Simulink must improve support object-oriented tools so that research community could go on to research supervisory control architectures, methodologies, and practical applications. A near future version Matlab/Simulink is announced to incorporate software structures. Structures will facilitate the application of these ideas to embed objects
176
J.LI. de la Rosa et al./Robotics and Autonomous Systems 21 (1997) 167-176
better. H o w e v e r , these structures are e x p e c t e d to bec o m e real objects in the next versions as required for AOP.
Acknowledgements This w o r k has b e e n partially f u n d e d by the TAP961114-C03-03 project Plataformas Integradas de C A D de Supervisi6n y Metodologfas o f C I C Y T p r o g r a m f r o m the Spanish g o v e r n m e n t .
References [1] N. Agell, J.C. Aguado and N. Piera, A qualitative mobil tracking, Current Trends in Qualitative Reasoning and Applications, Monograph CIMNE No 33 (1995) 102-107. [2] R.C. Arkin and K.S. Ali, Integration of reactive and telerobotic control in multi-agent robotic systems, Proc. 3th Int. Conf, on Simulation of Adaptive Behaviour. [3] T. Balch, G. Boone, T. Collins, H. Forbes, D. Mackenzie and J.C. Santamaria, Io, Ganymede and Callisto: A multiagent robot trash-collencting team, AI Magazine (16) (2) (1995) 39-51. [4] COORDINATION: Linguistic support for multiple cooperating agents, ESPRIT Project No 9102 in the Basic Research Action, Work Area: Basic Aspects of Multiple Computing Systems, Dates: 1.2.94-31.1.96. [5] J.Li. de la Rosa, J. Aguilar and I. Serra, Heuristics for Cooperation of Expert Systems. Application to Process Control, Ed. PIAR, 1994, ISBN 84-605-0275-9. [6] J.LI. de la Rosa, A. Oller et al., A comparison of fuzzy and qualitative control techniques, IEEE Int. Conf. on Control Applications~Int. Symp. on Intelligent Control~Int. Symp. on Computer-Aided Control System Design (ISBN 0-78032978-3) Dearborn, USA (1996) t39-144. [7] T. Finin et al,, DRAFT specification of the KQML (The knowledge query & manipulation language), unpublished draft, 1993. [8] P.J. Gmytrasiewicz, E.H. Durfee and D.K. Wehe, The utility of communication in coordinating intelligent agents, Proc. 9th National Conf. on Artificial Intelligence ( 1991 ). [9] J. Melendez, J.L1. de la Rosa and J. Aguilar, Embedding objects into Matlab/Simulink for process supervision, IEEE Control Systems Magazine (1997), to appear. [10] J. Melendez, J.LI. de la Rosa, J. Colomer, J. Vehi and C. Pous, Incrustaci6n de objetos en Simulink. Integraci6n de herramientas de ayuda al disefio de estructuras de supervisi6n, in: The H Congreso de usuarios de MATLAB, Madrid, Spain, September 1996. [11] Micro-Robot Soccer team Tournament, 9-12 November 1996, KAIST, Korea.
[12] C. Pous, A. Oiler, J. Vehf, J.LI. de la Rosa, Using real-time workshop in teaching control design techniques, Volume on Real Time Systems Education (IEEE Computer Society Press, Silver Spring, MD, 1996, ISBN 0-8186-7649-3) 153-158, August 1996. [ 13] Y. Shoham, Agent-oriented programming, Artificial Intelligence 60 (I 993) 51-92. [I4] M. Wooldridge and N.R. Jennings, Intelligent agents: Theory and practice, Knowledge Engineering Review, (1995), submitted.
Josep Lluis de la Rosa i Esteva obtained a degree in Sciences at the Autonomous University of Barcelona, Bellaterra (Spain) in 1989. In 1989 he joined Siemens-Nixdorf SoIL ware Development Center in Barcelona. In 1991 joined the Laboratoire d'Architecture et d'Analyse des Systbmes of the Centre National de la Rechercher Scientifique (LAAS-CNRS) in Toulouse (France), where he got the Ph.D. degree in Sciences by the Autonomous University of Barcelona in 1993. From 1994 he is in the University of Girona, adscribed to the Institute of Informatics and Application (IliA). His present research interests include artificial intelligence applied to supervisory systems and process control, and multi-agents systems. WWW: http://eia.udg.es/'-~peplluis. Albert Oiler i Pujol obtained a degree in Sciences at the Autonomous University of Barcelona, Bellaterra (Spain) in 1990. From 1994 he is in the University of Girona, adscribed to the Institute of Informatics and Application (IliA). His present research interests include artificial intelligence applied to supervisory systems and process control, and multi-agents systems. WWW: http://eia.udg.es/"~oller.
Josep PuyoI-Gruart obtained a degree in Electronic Engineering at the Polytechnic University of Catalonia, Barcelona (Spain) in 1985. From 1987 he has worked for the Spanish Council for Scientific Research (CSIC) at the Artificial Intelligence Research Institute (IliA). He got the Ph.D. degree in Computer Science by the Autonomous University of Barcelona in 1994. His present research interests include knowledge-based systems and multi-agents systems. WWW: http://www.iiia.csic.es/~puyol.
Josep Vehl obtained a degree in Sciences at the Autonomous University of Barcelona, Bellaterra (Spain) in 1987. Since September 1987 he has worked at the Technical University of Catalonia. Since 1991 he has been working at the University of Girona, adscribed to the Institute of Informatics and Applications (IIiA). His present research interests include interval methods applied to robust control parametric design and interval simulation.